Minty VI Gap in Variational Inequalities

• Minty VI gap is the discrepancy between Stampacchia (primal) and Minty (dual) solutions, highlighting a strict separation in variational inequality formulations.
• The gap function quantifies the worst-case dual violation, serving as a certificate for algorithmic proximity to Minty VI feasibility and guiding convergence analysis.
• Computational methods leveraging the Minty gap support robust convergence guarantees and feasibility certificates in settings like game theory, auction models, and non-monotone optimization.

A Minty VI gap refers to a fundamental discrepancy between solutions of the classical (Stampacchia) variational inequality (VI) and the Minty (dual) variational inequality (MVI). This concept is pivotal in variational inequality theory, computational optimization, game theory, and the convergence analysis of equilibrium learning algorithms—where the existence or nonexistence of an MVI solution, along with the magnitude of the Minty gap function, determines algorithmic tractability, solution quality, and problem complexity.

1. Classical and Minty Variational Inequality Formulations

Let $X$ be a (possibly infinite-dimensional) Banach or Hilbert space, $K \subset X$ a non-empty closed convex set, and $F: K \to X^*$ a (possibly nonlinear) mapping. The classical (Stampacchia) VI seeks $x^* \in K$ such that

$$\langle F(x^*), x - x^* \rangle \geq 0\quad \forall x \in K.$$

This condition is encoded by the Stampacchia (standard) VI gap function,

$$G_{\mathrm{VI}}(x) := \sup_{y \in K} \langle F(x), x - y \rangle.$$

A point $x^*$ solves VI if and only if $G_{\mathrm{VI}}(x^*) = 0$.

The Minty (dual) VI requires $x^M \in K$ such that

$$\langle F(y), y - x^M \rangle \geq 0\quad \forall y \in K.$$

The associated Minty gap function is

$$G_{\mathrm{M}}(x) := \sup_{y \in K} \langle F(y), y - x \rangle,$$

and $G_{\mathrm{M}}(x^M) = 0$ if and only if $x^M$ solves MVI. Every MVI solution is a VI solution, but the reverse does not generally hold; the difference between these two solution sets is captured by the Minty VI gap (Bichler et al., 2023, Lassonde, 2015, Anagnostides et al., 4 Apr 2025).

2. The Minty Condition and Monotonicity

A mapping $F$ is monotone if

$$\langle F(x) - F(y), x - y \rangle \geq 0\quad \forall x,y \in K.$$

When $F$ is monotone, the solution sets of VI and MVI coincide: the Minty condition (existence of an MVI solution) is both necessary and sufficient for monotonicity to fully characterize the problem. Minty's theorem asserts that, under monotonicity, every Minty VI solution is also a Stampacchia VI solution, and conversely. Lassonde established the converse: if the Minty condition holds for all compact convex $K$, then $F$ is monotone (Lassonde, 2015).

When $F$ is not monotone, the Minty VI may not admit a solution even when the Stampacchia VI does, and thus a Minty VI gap is present.

3. Diagnostic and Algorithmic Role of the Minty Gap

The Minty VI gap function serves as a certificate of proximity to MVI feasibility. It quantifies the worst-case dual violation: $$G_{\mathrm{M}}(x) = \sup_{y \in K} \langle F(y), y - x \rangle.$$ Certificates for Stampacchia VI use $G_{\mathrm{VI}}(x)$; for MVI, $G_{\mathrm{M}}(x)$. When $G_{\mathrm{M}}(x) \leq \epsilon$, algorithms (such as the extra-gradient ellipsoid method) can guarantee that $x$ is an approximate MVI solution, and—given further Lipschitz and boundedness assumptions—also nearly solves the Stampacchia VI (Anagnostides et al., 4 Apr 2025).

In optimization algorithms for VIs, convergence to an MVI (Minty-type) solution often ensures favorable properties: robustness to noise, global convergence, and performance guarantees in non-monotone contexts. In contrast, the presence of a nonzero Minty gap can obstruct these guarantees, limiting algorithmic power and necessitating alternative solution concepts (Zhao et al., 14 Oct 2025, Dey et al., 31 Jan 2026).

4. Minty Gap in Infinite-Dimensional and Game-Theoretic Settings

In the context of Bayesian auction games, equilibrium computation for first- and second-price mechanisms can be cast as infinite-dimensional VIs in suitably defined Banach spaces. The Minty gap sharply delineates behavior:

• For second-price auctions, the symmetric truthful-bidding equilibrium both solves the Stampacchia VI and the Minty VI: $G_{\mathrm{M}}(\beta^*) = 0$. Thus, the Minty condition holds, yielding tight convergence guarantees for learning algorithms.
• For first-price auctions, the Bayes-Nash equilibrium solves the Stampacchia VI but fails to solve the Minty VI; $G_{\mathrm{M}}(\beta^*) > 0$, and no Minty-type solution exists. This “Minty gap” precludes general convergence guarantees for extragradient or projection methods, although empirical convergence is sometimes observed (Bichler et al., 2023).

5. Computational Complexity and Algorithms for Minty Gap Problems

The Minty VI is computationally more demanding than the Stampacchia VI. Deciding whether an MVI solution exists (the Minty condition) is $\mathsf{coNP}$-complete, even for succinct normal-form games and two-player concave games via reduction from bilinear optimization (Anagnostides et al., 4 Apr 2025).

Nevertheless, under the Minty condition, (Anagnostides et al., 4 Apr 2025) provides a polynomial-time ellipsoid-based algorithm (ExtraGradientEllipsoid) for $\epsilon$-approximate VI solutions. The methodology leverages the Minty gap function for separation and achieves runtime polynomial in the dimension and $\log(1/\epsilon)$. If the Minty condition fails, the algorithm outputs a succinct certificate (in the form of a strict Expected VI solution) verifying infeasibility.

This algorithmic approach generalizes to (i) monotone VIs, (ii) global minimization of quasar-convex functions, and (iii) computation of Nash equilibria in harmonic games.

6. Smooth Minty Gap Functions, Error Bounds, and Relaxed Criteria

Smooth variants of the Minty gap, e.g., $$g_\mu(x) = \max_{y \in \mathcal{X}} \{ \langle F(x), x - y \rangle - \frac{1}{2\mu} \|x - y\|^2 \}$$, offer differentiability for algorithmic purposes and converge to the classical Minty gap as $\mu \to 0^+$ (Zhao et al., 14 Oct 2025). Error bound properties for these gap functions, both uniform and level-set types, are fundamental for establishing linear or global convergence of proximal-gradient and homotopy continuation schemes—crucially, even in the absence of an exact Minty solution.

Relaxed Minty gaps, combined with aggregated infeasibility measures, underpin ergodic $\mathcal{O}(1/K)$ convergence rates in misspecified or data-driven VI problems (Dey et al., 31 Jan 2026). The Minty gap remains effective as a diagnostic even off the feasible region, supporting algorithmic progress in practical settings.

7. Summary Table: VI, Minty VI, and Gap Functions

Formulation	Solution Set Definition	Gap Function
Stampacchia VI (Primal)	$\langle F(x^*), y - x^* \rangle \geq 0\ \forall y$	$$G_{\mathrm{VI}}(x) = \sup_y \langle F(x), x-y \rangle$$
Minty VI (Dual)	$\langle F(y), y - x^M \rangle \geq 0\ \forall y$	$$G_{\mathrm{M}}(x) = \sup_y \langle F(y), y-x \rangle$$
Smooth Minty Gap	$-$	$$g_\mu(x) = \max_y \langle F(x), x-y\rangle - \frac{1}{2\mu}\|x-y\|^2$$

A Minty VI gap, defined as $G_{\mathrm{M}}(x^*) > 0$ for $x^*$ solving the Stampacchia VI but not the Minty VI, signals a strict separation between primal and dual solution sets. The existence, size, and computational accessibility of this gap drive both theoretical and algorithmic frontiers in variational inequalities, optimization, and equilibrium computation (Bichler et al., 2023, Lassonde, 2015, Anagnostides et al., 4 Apr 2025, Zhao et al., 14 Oct 2025, Dey et al., 31 Jan 2026).